xref: /openbmc/linux/net/sched/sch_htb.c (revision 7211ec63)
1 /*
2  * net/sched/sch_htb.c	Hierarchical token bucket, feed tree version
3  *
4  *		This program is free software; you can redistribute it and/or
5  *		modify it under the terms of the GNU General Public License
6  *		as published by the Free Software Foundation; either version
7  *		2 of the License, or (at your option) any later version.
8  *
9  * Authors:	Martin Devera, <devik@cdi.cz>
10  *
11  * Credits (in time order) for older HTB versions:
12  *              Stef Coene <stef.coene@docum.org>
13  *			HTB support at LARTC mailing list
14  *		Ondrej Kraus, <krauso@barr.cz>
15  *			found missing INIT_QDISC(htb)
16  *		Vladimir Smelhaus, Aamer Akhter, Bert Hubert
17  *			helped a lot to locate nasty class stall bug
18  *		Andi Kleen, Jamal Hadi, Bert Hubert
19  *			code review and helpful comments on shaping
20  *		Tomasz Wrona, <tw@eter.tym.pl>
21  *			created test case so that I was able to fix nasty bug
22  *		Wilfried Weissmann
23  *			spotted bug in dequeue code and helped with fix
24  *		Jiri Fojtasek
25  *			fixed requeue routine
26  *		and many others. thanks.
27  */
28 #include <linux/module.h>
29 #include <linux/moduleparam.h>
30 #include <linux/types.h>
31 #include <linux/kernel.h>
32 #include <linux/string.h>
33 #include <linux/errno.h>
34 #include <linux/skbuff.h>
35 #include <linux/list.h>
36 #include <linux/compiler.h>
37 #include <linux/rbtree.h>
38 #include <linux/workqueue.h>
39 #include <linux/slab.h>
40 #include <net/netlink.h>
41 #include <net/sch_generic.h>
42 #include <net/pkt_sched.h>
43 #include <net/pkt_cls.h>
44 
45 /* HTB algorithm.
46     Author: devik@cdi.cz
47     ========================================================================
48     HTB is like TBF with multiple classes. It is also similar to CBQ because
49     it allows to assign priority to each class in hierarchy.
50     In fact it is another implementation of Floyd's formal sharing.
51 
52     Levels:
53     Each class is assigned level. Leaf has ALWAYS level 0 and root
54     classes have level TC_HTB_MAXDEPTH-1. Interior nodes has level
55     one less than their parent.
56 */
57 
58 static int htb_hysteresis __read_mostly = 0; /* whether to use mode hysteresis for speedup */
59 #define HTB_VER 0x30011		/* major must be matched with number suplied by TC as version */
60 
61 #if HTB_VER >> 16 != TC_HTB_PROTOVER
62 #error "Mismatched sch_htb.c and pkt_sch.h"
63 #endif
64 
65 /* Module parameter and sysfs export */
66 module_param    (htb_hysteresis, int, 0640);
67 MODULE_PARM_DESC(htb_hysteresis, "Hysteresis mode, less CPU load, less accurate");
68 
69 static int htb_rate_est = 0; /* htb classes have a default rate estimator */
70 module_param(htb_rate_est, int, 0640);
71 MODULE_PARM_DESC(htb_rate_est, "setup a default rate estimator (4sec 16sec) for htb classes");
72 
73 /* used internaly to keep status of single class */
74 enum htb_cmode {
75 	HTB_CANT_SEND,		/* class can't send and can't borrow */
76 	HTB_MAY_BORROW,		/* class can't send but may borrow */
77 	HTB_CAN_SEND		/* class can send */
78 };
79 
80 struct htb_prio {
81 	union {
82 		struct rb_root	row;
83 		struct rb_root	feed;
84 	};
85 	struct rb_node	*ptr;
86 	/* When class changes from state 1->2 and disconnects from
87 	 * parent's feed then we lost ptr value and start from the
88 	 * first child again. Here we store classid of the
89 	 * last valid ptr (used when ptr is NULL).
90 	 */
91 	u32		last_ptr_id;
92 };
93 
94 /* interior & leaf nodes; props specific to leaves are marked L:
95  * To reduce false sharing, place mostly read fields at beginning,
96  * and mostly written ones at the end.
97  */
98 struct htb_class {
99 	struct Qdisc_class_common common;
100 	struct psched_ratecfg	rate;
101 	struct psched_ratecfg	ceil;
102 	s64			buffer, cbuffer;/* token bucket depth/rate */
103 	s64			mbuffer;	/* max wait time */
104 	u32			prio;		/* these two are used only by leaves... */
105 	int			quantum;	/* but stored for parent-to-leaf return */
106 
107 	struct tcf_proto __rcu	*filter_list;	/* class attached filters */
108 	struct tcf_block	*block;
109 	int			filter_cnt;
110 
111 	int			level;		/* our level (see above) */
112 	unsigned int		children;
113 	struct htb_class	*parent;	/* parent class */
114 
115 	struct net_rate_estimator __rcu *rate_est;
116 
117 	/*
118 	 * Written often fields
119 	 */
120 	struct gnet_stats_basic_packed bstats;
121 	struct tc_htb_xstats	xstats;	/* our special stats */
122 
123 	/* token bucket parameters */
124 	s64			tokens, ctokens;/* current number of tokens */
125 	s64			t_c;		/* checkpoint time */
126 
127 	union {
128 		struct htb_class_leaf {
129 			struct list_head drop_list;
130 			int		deficit[TC_HTB_MAXDEPTH];
131 			struct Qdisc	*q;
132 		} leaf;
133 		struct htb_class_inner {
134 			struct htb_prio clprio[TC_HTB_NUMPRIO];
135 		} inner;
136 	} un;
137 	s64			pq_key;
138 
139 	int			prio_activity;	/* for which prios are we active */
140 	enum htb_cmode		cmode;		/* current mode of the class */
141 	struct rb_node		pq_node;	/* node for event queue */
142 	struct rb_node		node[TC_HTB_NUMPRIO];	/* node for self or feed tree */
143 
144 	unsigned int drops ____cacheline_aligned_in_smp;
145 };
146 
147 struct htb_level {
148 	struct rb_root	wait_pq;
149 	struct htb_prio hprio[TC_HTB_NUMPRIO];
150 };
151 
152 struct htb_sched {
153 	struct Qdisc_class_hash clhash;
154 	int			defcls;		/* class where unclassified flows go to */
155 	int			rate2quantum;	/* quant = rate / rate2quantum */
156 
157 	/* filters for qdisc itself */
158 	struct tcf_proto __rcu	*filter_list;
159 	struct tcf_block	*block;
160 
161 #define HTB_WARN_TOOMANYEVENTS	0x1
162 	unsigned int		warned;	/* only one warning */
163 	int			direct_qlen;
164 	struct work_struct	work;
165 
166 	/* non shaped skbs; let them go directly thru */
167 	struct qdisc_skb_head	direct_queue;
168 	long			direct_pkts;
169 
170 	struct qdisc_watchdog	watchdog;
171 
172 	s64			now;	/* cached dequeue time */
173 	struct list_head	drops[TC_HTB_NUMPRIO];/* active leaves (for drops) */
174 
175 	/* time of nearest event per level (row) */
176 	s64			near_ev_cache[TC_HTB_MAXDEPTH];
177 
178 	int			row_mask[TC_HTB_MAXDEPTH];
179 
180 	struct htb_level	hlevel[TC_HTB_MAXDEPTH];
181 };
182 
183 /* find class in global hash table using given handle */
184 static inline struct htb_class *htb_find(u32 handle, struct Qdisc *sch)
185 {
186 	struct htb_sched *q = qdisc_priv(sch);
187 	struct Qdisc_class_common *clc;
188 
189 	clc = qdisc_class_find(&q->clhash, handle);
190 	if (clc == NULL)
191 		return NULL;
192 	return container_of(clc, struct htb_class, common);
193 }
194 
195 static unsigned long htb_search(struct Qdisc *sch, u32 handle)
196 {
197 	return (unsigned long)htb_find(handle, sch);
198 }
199 /**
200  * htb_classify - classify a packet into class
201  *
202  * It returns NULL if the packet should be dropped or -1 if the packet
203  * should be passed directly thru. In all other cases leaf class is returned.
204  * We allow direct class selection by classid in priority. The we examine
205  * filters in qdisc and in inner nodes (if higher filter points to the inner
206  * node). If we end up with classid MAJOR:0 we enqueue the skb into special
207  * internal fifo (direct). These packets then go directly thru. If we still
208  * have no valid leaf we try to use MAJOR:default leaf. It still unsuccessful
209  * then finish and return direct queue.
210  */
211 #define HTB_DIRECT ((struct htb_class *)-1L)
212 
213 static struct htb_class *htb_classify(struct sk_buff *skb, struct Qdisc *sch,
214 				      int *qerr)
215 {
216 	struct htb_sched *q = qdisc_priv(sch);
217 	struct htb_class *cl;
218 	struct tcf_result res;
219 	struct tcf_proto *tcf;
220 	int result;
221 
222 	/* allow to select class by setting skb->priority to valid classid;
223 	 * note that nfmark can be used too by attaching filter fw with no
224 	 * rules in it
225 	 */
226 	if (skb->priority == sch->handle)
227 		return HTB_DIRECT;	/* X:0 (direct flow) selected */
228 	cl = htb_find(skb->priority, sch);
229 	if (cl) {
230 		if (cl->level == 0)
231 			return cl;
232 		/* Start with inner filter chain if a non-leaf class is selected */
233 		tcf = rcu_dereference_bh(cl->filter_list);
234 	} else {
235 		tcf = rcu_dereference_bh(q->filter_list);
236 	}
237 
238 	*qerr = NET_XMIT_SUCCESS | __NET_XMIT_BYPASS;
239 	while (tcf && (result = tcf_classify(skb, tcf, &res, false)) >= 0) {
240 #ifdef CONFIG_NET_CLS_ACT
241 		switch (result) {
242 		case TC_ACT_QUEUED:
243 		case TC_ACT_STOLEN:
244 		case TC_ACT_TRAP:
245 			*qerr = NET_XMIT_SUCCESS | __NET_XMIT_STOLEN;
246 		case TC_ACT_SHOT:
247 			return NULL;
248 		}
249 #endif
250 		cl = (void *)res.class;
251 		if (!cl) {
252 			if (res.classid == sch->handle)
253 				return HTB_DIRECT;	/* X:0 (direct flow) */
254 			cl = htb_find(res.classid, sch);
255 			if (!cl)
256 				break;	/* filter selected invalid classid */
257 		}
258 		if (!cl->level)
259 			return cl;	/* we hit leaf; return it */
260 
261 		/* we have got inner class; apply inner filter chain */
262 		tcf = rcu_dereference_bh(cl->filter_list);
263 	}
264 	/* classification failed; try to use default class */
265 	cl = htb_find(TC_H_MAKE(TC_H_MAJ(sch->handle), q->defcls), sch);
266 	if (!cl || cl->level)
267 		return HTB_DIRECT;	/* bad default .. this is safe bet */
268 	return cl;
269 }
270 
271 /**
272  * htb_add_to_id_tree - adds class to the round robin list
273  *
274  * Routine adds class to the list (actually tree) sorted by classid.
275  * Make sure that class is not already on such list for given prio.
276  */
277 static void htb_add_to_id_tree(struct rb_root *root,
278 			       struct htb_class *cl, int prio)
279 {
280 	struct rb_node **p = &root->rb_node, *parent = NULL;
281 
282 	while (*p) {
283 		struct htb_class *c;
284 		parent = *p;
285 		c = rb_entry(parent, struct htb_class, node[prio]);
286 
287 		if (cl->common.classid > c->common.classid)
288 			p = &parent->rb_right;
289 		else
290 			p = &parent->rb_left;
291 	}
292 	rb_link_node(&cl->node[prio], parent, p);
293 	rb_insert_color(&cl->node[prio], root);
294 }
295 
296 /**
297  * htb_add_to_wait_tree - adds class to the event queue with delay
298  *
299  * The class is added to priority event queue to indicate that class will
300  * change its mode in cl->pq_key microseconds. Make sure that class is not
301  * already in the queue.
302  */
303 static void htb_add_to_wait_tree(struct htb_sched *q,
304 				 struct htb_class *cl, s64 delay)
305 {
306 	struct rb_node **p = &q->hlevel[cl->level].wait_pq.rb_node, *parent = NULL;
307 
308 	cl->pq_key = q->now + delay;
309 	if (cl->pq_key == q->now)
310 		cl->pq_key++;
311 
312 	/* update the nearest event cache */
313 	if (q->near_ev_cache[cl->level] > cl->pq_key)
314 		q->near_ev_cache[cl->level] = cl->pq_key;
315 
316 	while (*p) {
317 		struct htb_class *c;
318 		parent = *p;
319 		c = rb_entry(parent, struct htb_class, pq_node);
320 		if (cl->pq_key >= c->pq_key)
321 			p = &parent->rb_right;
322 		else
323 			p = &parent->rb_left;
324 	}
325 	rb_link_node(&cl->pq_node, parent, p);
326 	rb_insert_color(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
327 }
328 
329 /**
330  * htb_next_rb_node - finds next node in binary tree
331  *
332  * When we are past last key we return NULL.
333  * Average complexity is 2 steps per call.
334  */
335 static inline void htb_next_rb_node(struct rb_node **n)
336 {
337 	*n = rb_next(*n);
338 }
339 
340 /**
341  * htb_add_class_to_row - add class to its row
342  *
343  * The class is added to row at priorities marked in mask.
344  * It does nothing if mask == 0.
345  */
346 static inline void htb_add_class_to_row(struct htb_sched *q,
347 					struct htb_class *cl, int mask)
348 {
349 	q->row_mask[cl->level] |= mask;
350 	while (mask) {
351 		int prio = ffz(~mask);
352 		mask &= ~(1 << prio);
353 		htb_add_to_id_tree(&q->hlevel[cl->level].hprio[prio].row, cl, prio);
354 	}
355 }
356 
357 /* If this triggers, it is a bug in this code, but it need not be fatal */
358 static void htb_safe_rb_erase(struct rb_node *rb, struct rb_root *root)
359 {
360 	if (RB_EMPTY_NODE(rb)) {
361 		WARN_ON(1);
362 	} else {
363 		rb_erase(rb, root);
364 		RB_CLEAR_NODE(rb);
365 	}
366 }
367 
368 
369 /**
370  * htb_remove_class_from_row - removes class from its row
371  *
372  * The class is removed from row at priorities marked in mask.
373  * It does nothing if mask == 0.
374  */
375 static inline void htb_remove_class_from_row(struct htb_sched *q,
376 						 struct htb_class *cl, int mask)
377 {
378 	int m = 0;
379 	struct htb_level *hlevel = &q->hlevel[cl->level];
380 
381 	while (mask) {
382 		int prio = ffz(~mask);
383 		struct htb_prio *hprio = &hlevel->hprio[prio];
384 
385 		mask &= ~(1 << prio);
386 		if (hprio->ptr == cl->node + prio)
387 			htb_next_rb_node(&hprio->ptr);
388 
389 		htb_safe_rb_erase(cl->node + prio, &hprio->row);
390 		if (!hprio->row.rb_node)
391 			m |= 1 << prio;
392 	}
393 	q->row_mask[cl->level] &= ~m;
394 }
395 
396 /**
397  * htb_activate_prios - creates active classe's feed chain
398  *
399  * The class is connected to ancestors and/or appropriate rows
400  * for priorities it is participating on. cl->cmode must be new
401  * (activated) mode. It does nothing if cl->prio_activity == 0.
402  */
403 static void htb_activate_prios(struct htb_sched *q, struct htb_class *cl)
404 {
405 	struct htb_class *p = cl->parent;
406 	long m, mask = cl->prio_activity;
407 
408 	while (cl->cmode == HTB_MAY_BORROW && p && mask) {
409 		m = mask;
410 		while (m) {
411 			int prio = ffz(~m);
412 			m &= ~(1 << prio);
413 
414 			if (p->un.inner.clprio[prio].feed.rb_node)
415 				/* parent already has its feed in use so that
416 				 * reset bit in mask as parent is already ok
417 				 */
418 				mask &= ~(1 << prio);
419 
420 			htb_add_to_id_tree(&p->un.inner.clprio[prio].feed, cl, prio);
421 		}
422 		p->prio_activity |= mask;
423 		cl = p;
424 		p = cl->parent;
425 
426 	}
427 	if (cl->cmode == HTB_CAN_SEND && mask)
428 		htb_add_class_to_row(q, cl, mask);
429 }
430 
431 /**
432  * htb_deactivate_prios - remove class from feed chain
433  *
434  * cl->cmode must represent old mode (before deactivation). It does
435  * nothing if cl->prio_activity == 0. Class is removed from all feed
436  * chains and rows.
437  */
438 static void htb_deactivate_prios(struct htb_sched *q, struct htb_class *cl)
439 {
440 	struct htb_class *p = cl->parent;
441 	long m, mask = cl->prio_activity;
442 
443 	while (cl->cmode == HTB_MAY_BORROW && p && mask) {
444 		m = mask;
445 		mask = 0;
446 		while (m) {
447 			int prio = ffz(~m);
448 			m &= ~(1 << prio);
449 
450 			if (p->un.inner.clprio[prio].ptr == cl->node + prio) {
451 				/* we are removing child which is pointed to from
452 				 * parent feed - forget the pointer but remember
453 				 * classid
454 				 */
455 				p->un.inner.clprio[prio].last_ptr_id = cl->common.classid;
456 				p->un.inner.clprio[prio].ptr = NULL;
457 			}
458 
459 			htb_safe_rb_erase(cl->node + prio,
460 					  &p->un.inner.clprio[prio].feed);
461 
462 			if (!p->un.inner.clprio[prio].feed.rb_node)
463 				mask |= 1 << prio;
464 		}
465 
466 		p->prio_activity &= ~mask;
467 		cl = p;
468 		p = cl->parent;
469 
470 	}
471 	if (cl->cmode == HTB_CAN_SEND && mask)
472 		htb_remove_class_from_row(q, cl, mask);
473 }
474 
475 static inline s64 htb_lowater(const struct htb_class *cl)
476 {
477 	if (htb_hysteresis)
478 		return cl->cmode != HTB_CANT_SEND ? -cl->cbuffer : 0;
479 	else
480 		return 0;
481 }
482 static inline s64 htb_hiwater(const struct htb_class *cl)
483 {
484 	if (htb_hysteresis)
485 		return cl->cmode == HTB_CAN_SEND ? -cl->buffer : 0;
486 	else
487 		return 0;
488 }
489 
490 
491 /**
492  * htb_class_mode - computes and returns current class mode
493  *
494  * It computes cl's mode at time cl->t_c+diff and returns it. If mode
495  * is not HTB_CAN_SEND then cl->pq_key is updated to time difference
496  * from now to time when cl will change its state.
497  * Also it is worth to note that class mode doesn't change simply
498  * at cl->{c,}tokens == 0 but there can rather be hysteresis of
499  * 0 .. -cl->{c,}buffer range. It is meant to limit number of
500  * mode transitions per time unit. The speed gain is about 1/6.
501  */
502 static inline enum htb_cmode
503 htb_class_mode(struct htb_class *cl, s64 *diff)
504 {
505 	s64 toks;
506 
507 	if ((toks = (cl->ctokens + *diff)) < htb_lowater(cl)) {
508 		*diff = -toks;
509 		return HTB_CANT_SEND;
510 	}
511 
512 	if ((toks = (cl->tokens + *diff)) >= htb_hiwater(cl))
513 		return HTB_CAN_SEND;
514 
515 	*diff = -toks;
516 	return HTB_MAY_BORROW;
517 }
518 
519 /**
520  * htb_change_class_mode - changes classe's mode
521  *
522  * This should be the only way how to change classe's mode under normal
523  * cirsumstances. Routine will update feed lists linkage, change mode
524  * and add class to the wait event queue if appropriate. New mode should
525  * be different from old one and cl->pq_key has to be valid if changing
526  * to mode other than HTB_CAN_SEND (see htb_add_to_wait_tree).
527  */
528 static void
529 htb_change_class_mode(struct htb_sched *q, struct htb_class *cl, s64 *diff)
530 {
531 	enum htb_cmode new_mode = htb_class_mode(cl, diff);
532 
533 	if (new_mode == cl->cmode)
534 		return;
535 
536 	if (cl->prio_activity) {	/* not necessary: speed optimization */
537 		if (cl->cmode != HTB_CANT_SEND)
538 			htb_deactivate_prios(q, cl);
539 		cl->cmode = new_mode;
540 		if (new_mode != HTB_CANT_SEND)
541 			htb_activate_prios(q, cl);
542 	} else
543 		cl->cmode = new_mode;
544 }
545 
546 /**
547  * htb_activate - inserts leaf cl into appropriate active feeds
548  *
549  * Routine learns (new) priority of leaf and activates feed chain
550  * for the prio. It can be called on already active leaf safely.
551  * It also adds leaf into droplist.
552  */
553 static inline void htb_activate(struct htb_sched *q, struct htb_class *cl)
554 {
555 	WARN_ON(cl->level || !cl->un.leaf.q || !cl->un.leaf.q->q.qlen);
556 
557 	if (!cl->prio_activity) {
558 		cl->prio_activity = 1 << cl->prio;
559 		htb_activate_prios(q, cl);
560 		list_add_tail(&cl->un.leaf.drop_list,
561 			      q->drops + cl->prio);
562 	}
563 }
564 
565 /**
566  * htb_deactivate - remove leaf cl from active feeds
567  *
568  * Make sure that leaf is active. In the other words it can't be called
569  * with non-active leaf. It also removes class from the drop list.
570  */
571 static inline void htb_deactivate(struct htb_sched *q, struct htb_class *cl)
572 {
573 	WARN_ON(!cl->prio_activity);
574 
575 	htb_deactivate_prios(q, cl);
576 	cl->prio_activity = 0;
577 	list_del_init(&cl->un.leaf.drop_list);
578 }
579 
580 static void htb_enqueue_tail(struct sk_buff *skb, struct Qdisc *sch,
581 			     struct qdisc_skb_head *qh)
582 {
583 	struct sk_buff *last = qh->tail;
584 
585 	if (last) {
586 		skb->next = NULL;
587 		last->next = skb;
588 		qh->tail = skb;
589 	} else {
590 		qh->tail = skb;
591 		qh->head = skb;
592 	}
593 	qh->qlen++;
594 }
595 
596 static int htb_enqueue(struct sk_buff *skb, struct Qdisc *sch,
597 		       struct sk_buff **to_free)
598 {
599 	int uninitialized_var(ret);
600 	struct htb_sched *q = qdisc_priv(sch);
601 	struct htb_class *cl = htb_classify(skb, sch, &ret);
602 
603 	if (cl == HTB_DIRECT) {
604 		/* enqueue to helper queue */
605 		if (q->direct_queue.qlen < q->direct_qlen) {
606 			htb_enqueue_tail(skb, sch, &q->direct_queue);
607 			q->direct_pkts++;
608 		} else {
609 			return qdisc_drop(skb, sch, to_free);
610 		}
611 #ifdef CONFIG_NET_CLS_ACT
612 	} else if (!cl) {
613 		if (ret & __NET_XMIT_BYPASS)
614 			qdisc_qstats_drop(sch);
615 		__qdisc_drop(skb, to_free);
616 		return ret;
617 #endif
618 	} else if ((ret = qdisc_enqueue(skb, cl->un.leaf.q,
619 					to_free)) != NET_XMIT_SUCCESS) {
620 		if (net_xmit_drop_count(ret)) {
621 			qdisc_qstats_drop(sch);
622 			cl->drops++;
623 		}
624 		return ret;
625 	} else {
626 		htb_activate(q, cl);
627 	}
628 
629 	qdisc_qstats_backlog_inc(sch, skb);
630 	sch->q.qlen++;
631 	return NET_XMIT_SUCCESS;
632 }
633 
634 static inline void htb_accnt_tokens(struct htb_class *cl, int bytes, s64 diff)
635 {
636 	s64 toks = diff + cl->tokens;
637 
638 	if (toks > cl->buffer)
639 		toks = cl->buffer;
640 	toks -= (s64) psched_l2t_ns(&cl->rate, bytes);
641 	if (toks <= -cl->mbuffer)
642 		toks = 1 - cl->mbuffer;
643 
644 	cl->tokens = toks;
645 }
646 
647 static inline void htb_accnt_ctokens(struct htb_class *cl, int bytes, s64 diff)
648 {
649 	s64 toks = diff + cl->ctokens;
650 
651 	if (toks > cl->cbuffer)
652 		toks = cl->cbuffer;
653 	toks -= (s64) psched_l2t_ns(&cl->ceil, bytes);
654 	if (toks <= -cl->mbuffer)
655 		toks = 1 - cl->mbuffer;
656 
657 	cl->ctokens = toks;
658 }
659 
660 /**
661  * htb_charge_class - charges amount "bytes" to leaf and ancestors
662  *
663  * Routine assumes that packet "bytes" long was dequeued from leaf cl
664  * borrowing from "level". It accounts bytes to ceil leaky bucket for
665  * leaf and all ancestors and to rate bucket for ancestors at levels
666  * "level" and higher. It also handles possible change of mode resulting
667  * from the update. Note that mode can also increase here (MAY_BORROW to
668  * CAN_SEND) because we can use more precise clock that event queue here.
669  * In such case we remove class from event queue first.
670  */
671 static void htb_charge_class(struct htb_sched *q, struct htb_class *cl,
672 			     int level, struct sk_buff *skb)
673 {
674 	int bytes = qdisc_pkt_len(skb);
675 	enum htb_cmode old_mode;
676 	s64 diff;
677 
678 	while (cl) {
679 		diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
680 		if (cl->level >= level) {
681 			if (cl->level == level)
682 				cl->xstats.lends++;
683 			htb_accnt_tokens(cl, bytes, diff);
684 		} else {
685 			cl->xstats.borrows++;
686 			cl->tokens += diff;	/* we moved t_c; update tokens */
687 		}
688 		htb_accnt_ctokens(cl, bytes, diff);
689 		cl->t_c = q->now;
690 
691 		old_mode = cl->cmode;
692 		diff = 0;
693 		htb_change_class_mode(q, cl, &diff);
694 		if (old_mode != cl->cmode) {
695 			if (old_mode != HTB_CAN_SEND)
696 				htb_safe_rb_erase(&cl->pq_node, &q->hlevel[cl->level].wait_pq);
697 			if (cl->cmode != HTB_CAN_SEND)
698 				htb_add_to_wait_tree(q, cl, diff);
699 		}
700 
701 		/* update basic stats except for leaves which are already updated */
702 		if (cl->level)
703 			bstats_update(&cl->bstats, skb);
704 
705 		cl = cl->parent;
706 	}
707 }
708 
709 /**
710  * htb_do_events - make mode changes to classes at the level
711  *
712  * Scans event queue for pending events and applies them. Returns time of
713  * next pending event (0 for no event in pq, q->now for too many events).
714  * Note: Applied are events whose have cl->pq_key <= q->now.
715  */
716 static s64 htb_do_events(struct htb_sched *q, const int level,
717 			 unsigned long start)
718 {
719 	/* don't run for longer than 2 jiffies; 2 is used instead of
720 	 * 1 to simplify things when jiffy is going to be incremented
721 	 * too soon
722 	 */
723 	unsigned long stop_at = start + 2;
724 	struct rb_root *wait_pq = &q->hlevel[level].wait_pq;
725 
726 	while (time_before(jiffies, stop_at)) {
727 		struct htb_class *cl;
728 		s64 diff;
729 		struct rb_node *p = rb_first(wait_pq);
730 
731 		if (!p)
732 			return 0;
733 
734 		cl = rb_entry(p, struct htb_class, pq_node);
735 		if (cl->pq_key > q->now)
736 			return cl->pq_key;
737 
738 		htb_safe_rb_erase(p, wait_pq);
739 		diff = min_t(s64, q->now - cl->t_c, cl->mbuffer);
740 		htb_change_class_mode(q, cl, &diff);
741 		if (cl->cmode != HTB_CAN_SEND)
742 			htb_add_to_wait_tree(q, cl, diff);
743 	}
744 
745 	/* too much load - let's continue after a break for scheduling */
746 	if (!(q->warned & HTB_WARN_TOOMANYEVENTS)) {
747 		pr_warn("htb: too many events!\n");
748 		q->warned |= HTB_WARN_TOOMANYEVENTS;
749 	}
750 
751 	return q->now;
752 }
753 
754 /* Returns class->node+prio from id-tree where classe's id is >= id. NULL
755  * is no such one exists.
756  */
757 static struct rb_node *htb_id_find_next_upper(int prio, struct rb_node *n,
758 					      u32 id)
759 {
760 	struct rb_node *r = NULL;
761 	while (n) {
762 		struct htb_class *cl =
763 		    rb_entry(n, struct htb_class, node[prio]);
764 
765 		if (id > cl->common.classid) {
766 			n = n->rb_right;
767 		} else if (id < cl->common.classid) {
768 			r = n;
769 			n = n->rb_left;
770 		} else {
771 			return n;
772 		}
773 	}
774 	return r;
775 }
776 
777 /**
778  * htb_lookup_leaf - returns next leaf class in DRR order
779  *
780  * Find leaf where current feed pointers points to.
781  */
782 static struct htb_class *htb_lookup_leaf(struct htb_prio *hprio, const int prio)
783 {
784 	int i;
785 	struct {
786 		struct rb_node *root;
787 		struct rb_node **pptr;
788 		u32 *pid;
789 	} stk[TC_HTB_MAXDEPTH], *sp = stk;
790 
791 	BUG_ON(!hprio->row.rb_node);
792 	sp->root = hprio->row.rb_node;
793 	sp->pptr = &hprio->ptr;
794 	sp->pid = &hprio->last_ptr_id;
795 
796 	for (i = 0; i < 65535; i++) {
797 		if (!*sp->pptr && *sp->pid) {
798 			/* ptr was invalidated but id is valid - try to recover
799 			 * the original or next ptr
800 			 */
801 			*sp->pptr =
802 			    htb_id_find_next_upper(prio, sp->root, *sp->pid);
803 		}
804 		*sp->pid = 0;	/* ptr is valid now so that remove this hint as it
805 				 * can become out of date quickly
806 				 */
807 		if (!*sp->pptr) {	/* we are at right end; rewind & go up */
808 			*sp->pptr = sp->root;
809 			while ((*sp->pptr)->rb_left)
810 				*sp->pptr = (*sp->pptr)->rb_left;
811 			if (sp > stk) {
812 				sp--;
813 				if (!*sp->pptr) {
814 					WARN_ON(1);
815 					return NULL;
816 				}
817 				htb_next_rb_node(sp->pptr);
818 			}
819 		} else {
820 			struct htb_class *cl;
821 			struct htb_prio *clp;
822 
823 			cl = rb_entry(*sp->pptr, struct htb_class, node[prio]);
824 			if (!cl->level)
825 				return cl;
826 			clp = &cl->un.inner.clprio[prio];
827 			(++sp)->root = clp->feed.rb_node;
828 			sp->pptr = &clp->ptr;
829 			sp->pid = &clp->last_ptr_id;
830 		}
831 	}
832 	WARN_ON(1);
833 	return NULL;
834 }
835 
836 /* dequeues packet at given priority and level; call only if
837  * you are sure that there is active class at prio/level
838  */
839 static struct sk_buff *htb_dequeue_tree(struct htb_sched *q, const int prio,
840 					const int level)
841 {
842 	struct sk_buff *skb = NULL;
843 	struct htb_class *cl, *start;
844 	struct htb_level *hlevel = &q->hlevel[level];
845 	struct htb_prio *hprio = &hlevel->hprio[prio];
846 
847 	/* look initial class up in the row */
848 	start = cl = htb_lookup_leaf(hprio, prio);
849 
850 	do {
851 next:
852 		if (unlikely(!cl))
853 			return NULL;
854 
855 		/* class can be empty - it is unlikely but can be true if leaf
856 		 * qdisc drops packets in enqueue routine or if someone used
857 		 * graft operation on the leaf since last dequeue;
858 		 * simply deactivate and skip such class
859 		 */
860 		if (unlikely(cl->un.leaf.q->q.qlen == 0)) {
861 			struct htb_class *next;
862 			htb_deactivate(q, cl);
863 
864 			/* row/level might become empty */
865 			if ((q->row_mask[level] & (1 << prio)) == 0)
866 				return NULL;
867 
868 			next = htb_lookup_leaf(hprio, prio);
869 
870 			if (cl == start)	/* fix start if we just deleted it */
871 				start = next;
872 			cl = next;
873 			goto next;
874 		}
875 
876 		skb = cl->un.leaf.q->dequeue(cl->un.leaf.q);
877 		if (likely(skb != NULL))
878 			break;
879 
880 		qdisc_warn_nonwc("htb", cl->un.leaf.q);
881 		htb_next_rb_node(level ? &cl->parent->un.inner.clprio[prio].ptr:
882 					 &q->hlevel[0].hprio[prio].ptr);
883 		cl = htb_lookup_leaf(hprio, prio);
884 
885 	} while (cl != start);
886 
887 	if (likely(skb != NULL)) {
888 		bstats_update(&cl->bstats, skb);
889 		cl->un.leaf.deficit[level] -= qdisc_pkt_len(skb);
890 		if (cl->un.leaf.deficit[level] < 0) {
891 			cl->un.leaf.deficit[level] += cl->quantum;
892 			htb_next_rb_node(level ? &cl->parent->un.inner.clprio[prio].ptr :
893 						 &q->hlevel[0].hprio[prio].ptr);
894 		}
895 		/* this used to be after charge_class but this constelation
896 		 * gives us slightly better performance
897 		 */
898 		if (!cl->un.leaf.q->q.qlen)
899 			htb_deactivate(q, cl);
900 		htb_charge_class(q, cl, level, skb);
901 	}
902 	return skb;
903 }
904 
905 static struct sk_buff *htb_dequeue(struct Qdisc *sch)
906 {
907 	struct sk_buff *skb;
908 	struct htb_sched *q = qdisc_priv(sch);
909 	int level;
910 	s64 next_event;
911 	unsigned long start_at;
912 
913 	/* try to dequeue direct packets as high prio (!) to minimize cpu work */
914 	skb = __qdisc_dequeue_head(&q->direct_queue);
915 	if (skb != NULL) {
916 ok:
917 		qdisc_bstats_update(sch, skb);
918 		qdisc_qstats_backlog_dec(sch, skb);
919 		sch->q.qlen--;
920 		return skb;
921 	}
922 
923 	if (!sch->q.qlen)
924 		goto fin;
925 	q->now = ktime_get_ns();
926 	start_at = jiffies;
927 
928 	next_event = q->now + 5LLU * NSEC_PER_SEC;
929 
930 	for (level = 0; level < TC_HTB_MAXDEPTH; level++) {
931 		/* common case optimization - skip event handler quickly */
932 		int m;
933 		s64 event = q->near_ev_cache[level];
934 
935 		if (q->now >= event) {
936 			event = htb_do_events(q, level, start_at);
937 			if (!event)
938 				event = q->now + NSEC_PER_SEC;
939 			q->near_ev_cache[level] = event;
940 		}
941 
942 		if (next_event > event)
943 			next_event = event;
944 
945 		m = ~q->row_mask[level];
946 		while (m != (int)(-1)) {
947 			int prio = ffz(m);
948 
949 			m |= 1 << prio;
950 			skb = htb_dequeue_tree(q, prio, level);
951 			if (likely(skb != NULL))
952 				goto ok;
953 		}
954 	}
955 	qdisc_qstats_overlimit(sch);
956 	if (likely(next_event > q->now))
957 		qdisc_watchdog_schedule_ns(&q->watchdog, next_event);
958 	else
959 		schedule_work(&q->work);
960 fin:
961 	return skb;
962 }
963 
964 /* reset all classes */
965 /* always caled under BH & queue lock */
966 static void htb_reset(struct Qdisc *sch)
967 {
968 	struct htb_sched *q = qdisc_priv(sch);
969 	struct htb_class *cl;
970 	unsigned int i;
971 
972 	for (i = 0; i < q->clhash.hashsize; i++) {
973 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
974 			if (cl->level)
975 				memset(&cl->un.inner, 0, sizeof(cl->un.inner));
976 			else {
977 				if (cl->un.leaf.q)
978 					qdisc_reset(cl->un.leaf.q);
979 				INIT_LIST_HEAD(&cl->un.leaf.drop_list);
980 			}
981 			cl->prio_activity = 0;
982 			cl->cmode = HTB_CAN_SEND;
983 		}
984 	}
985 	qdisc_watchdog_cancel(&q->watchdog);
986 	__qdisc_reset_queue(&q->direct_queue);
987 	sch->q.qlen = 0;
988 	sch->qstats.backlog = 0;
989 	memset(q->hlevel, 0, sizeof(q->hlevel));
990 	memset(q->row_mask, 0, sizeof(q->row_mask));
991 	for (i = 0; i < TC_HTB_NUMPRIO; i++)
992 		INIT_LIST_HEAD(q->drops + i);
993 }
994 
995 static const struct nla_policy htb_policy[TCA_HTB_MAX + 1] = {
996 	[TCA_HTB_PARMS]	= { .len = sizeof(struct tc_htb_opt) },
997 	[TCA_HTB_INIT]	= { .len = sizeof(struct tc_htb_glob) },
998 	[TCA_HTB_CTAB]	= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
999 	[TCA_HTB_RTAB]	= { .type = NLA_BINARY, .len = TC_RTAB_SIZE },
1000 	[TCA_HTB_DIRECT_QLEN] = { .type = NLA_U32 },
1001 	[TCA_HTB_RATE64] = { .type = NLA_U64 },
1002 	[TCA_HTB_CEIL64] = { .type = NLA_U64 },
1003 };
1004 
1005 static void htb_work_func(struct work_struct *work)
1006 {
1007 	struct htb_sched *q = container_of(work, struct htb_sched, work);
1008 	struct Qdisc *sch = q->watchdog.qdisc;
1009 
1010 	rcu_read_lock();
1011 	__netif_schedule(qdisc_root(sch));
1012 	rcu_read_unlock();
1013 }
1014 
1015 static int htb_init(struct Qdisc *sch, struct nlattr *opt)
1016 {
1017 	struct htb_sched *q = qdisc_priv(sch);
1018 	struct nlattr *tb[TCA_HTB_MAX + 1];
1019 	struct tc_htb_glob *gopt;
1020 	int err;
1021 	int i;
1022 
1023 	qdisc_watchdog_init(&q->watchdog, sch);
1024 	INIT_WORK(&q->work, htb_work_func);
1025 
1026 	if (!opt)
1027 		return -EINVAL;
1028 
1029 	err = tcf_block_get(&q->block, &q->filter_list);
1030 	if (err)
1031 		return err;
1032 
1033 	err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy, NULL);
1034 	if (err < 0)
1035 		return err;
1036 
1037 	if (!tb[TCA_HTB_INIT])
1038 		return -EINVAL;
1039 
1040 	gopt = nla_data(tb[TCA_HTB_INIT]);
1041 	if (gopt->version != HTB_VER >> 16)
1042 		return -EINVAL;
1043 
1044 	err = qdisc_class_hash_init(&q->clhash);
1045 	if (err < 0)
1046 		return err;
1047 	for (i = 0; i < TC_HTB_NUMPRIO; i++)
1048 		INIT_LIST_HEAD(q->drops + i);
1049 
1050 	qdisc_skb_head_init(&q->direct_queue);
1051 
1052 	if (tb[TCA_HTB_DIRECT_QLEN])
1053 		q->direct_qlen = nla_get_u32(tb[TCA_HTB_DIRECT_QLEN]);
1054 	else
1055 		q->direct_qlen = qdisc_dev(sch)->tx_queue_len;
1056 
1057 	if ((q->rate2quantum = gopt->rate2quantum) < 1)
1058 		q->rate2quantum = 1;
1059 	q->defcls = gopt->defcls;
1060 
1061 	return 0;
1062 }
1063 
1064 static int htb_dump(struct Qdisc *sch, struct sk_buff *skb)
1065 {
1066 	struct htb_sched *q = qdisc_priv(sch);
1067 	struct nlattr *nest;
1068 	struct tc_htb_glob gopt;
1069 
1070 	/* Its safe to not acquire qdisc lock. As we hold RTNL,
1071 	 * no change can happen on the qdisc parameters.
1072 	 */
1073 
1074 	gopt.direct_pkts = q->direct_pkts;
1075 	gopt.version = HTB_VER;
1076 	gopt.rate2quantum = q->rate2quantum;
1077 	gopt.defcls = q->defcls;
1078 	gopt.debug = 0;
1079 
1080 	nest = nla_nest_start(skb, TCA_OPTIONS);
1081 	if (nest == NULL)
1082 		goto nla_put_failure;
1083 	if (nla_put(skb, TCA_HTB_INIT, sizeof(gopt), &gopt) ||
1084 	    nla_put_u32(skb, TCA_HTB_DIRECT_QLEN, q->direct_qlen))
1085 		goto nla_put_failure;
1086 
1087 	return nla_nest_end(skb, nest);
1088 
1089 nla_put_failure:
1090 	nla_nest_cancel(skb, nest);
1091 	return -1;
1092 }
1093 
1094 static int htb_dump_class(struct Qdisc *sch, unsigned long arg,
1095 			  struct sk_buff *skb, struct tcmsg *tcm)
1096 {
1097 	struct htb_class *cl = (struct htb_class *)arg;
1098 	struct nlattr *nest;
1099 	struct tc_htb_opt opt;
1100 
1101 	/* Its safe to not acquire qdisc lock. As we hold RTNL,
1102 	 * no change can happen on the class parameters.
1103 	 */
1104 	tcm->tcm_parent = cl->parent ? cl->parent->common.classid : TC_H_ROOT;
1105 	tcm->tcm_handle = cl->common.classid;
1106 	if (!cl->level && cl->un.leaf.q)
1107 		tcm->tcm_info = cl->un.leaf.q->handle;
1108 
1109 	nest = nla_nest_start(skb, TCA_OPTIONS);
1110 	if (nest == NULL)
1111 		goto nla_put_failure;
1112 
1113 	memset(&opt, 0, sizeof(opt));
1114 
1115 	psched_ratecfg_getrate(&opt.rate, &cl->rate);
1116 	opt.buffer = PSCHED_NS2TICKS(cl->buffer);
1117 	psched_ratecfg_getrate(&opt.ceil, &cl->ceil);
1118 	opt.cbuffer = PSCHED_NS2TICKS(cl->cbuffer);
1119 	opt.quantum = cl->quantum;
1120 	opt.prio = cl->prio;
1121 	opt.level = cl->level;
1122 	if (nla_put(skb, TCA_HTB_PARMS, sizeof(opt), &opt))
1123 		goto nla_put_failure;
1124 	if ((cl->rate.rate_bytes_ps >= (1ULL << 32)) &&
1125 	    nla_put_u64_64bit(skb, TCA_HTB_RATE64, cl->rate.rate_bytes_ps,
1126 			      TCA_HTB_PAD))
1127 		goto nla_put_failure;
1128 	if ((cl->ceil.rate_bytes_ps >= (1ULL << 32)) &&
1129 	    nla_put_u64_64bit(skb, TCA_HTB_CEIL64, cl->ceil.rate_bytes_ps,
1130 			      TCA_HTB_PAD))
1131 		goto nla_put_failure;
1132 
1133 	return nla_nest_end(skb, nest);
1134 
1135 nla_put_failure:
1136 	nla_nest_cancel(skb, nest);
1137 	return -1;
1138 }
1139 
1140 static int
1141 htb_dump_class_stats(struct Qdisc *sch, unsigned long arg, struct gnet_dump *d)
1142 {
1143 	struct htb_class *cl = (struct htb_class *)arg;
1144 	struct gnet_stats_queue qs = {
1145 		.drops = cl->drops,
1146 	};
1147 	__u32 qlen = 0;
1148 
1149 	if (!cl->level && cl->un.leaf.q) {
1150 		qlen = cl->un.leaf.q->q.qlen;
1151 		qs.backlog = cl->un.leaf.q->qstats.backlog;
1152 	}
1153 	cl->xstats.tokens = clamp_t(s64, PSCHED_NS2TICKS(cl->tokens),
1154 				    INT_MIN, INT_MAX);
1155 	cl->xstats.ctokens = clamp_t(s64, PSCHED_NS2TICKS(cl->ctokens),
1156 				     INT_MIN, INT_MAX);
1157 
1158 	if (gnet_stats_copy_basic(qdisc_root_sleeping_running(sch),
1159 				  d, NULL, &cl->bstats) < 0 ||
1160 	    gnet_stats_copy_rate_est(d, &cl->rate_est) < 0 ||
1161 	    gnet_stats_copy_queue(d, NULL, &qs, qlen) < 0)
1162 		return -1;
1163 
1164 	return gnet_stats_copy_app(d, &cl->xstats, sizeof(cl->xstats));
1165 }
1166 
1167 static int htb_graft(struct Qdisc *sch, unsigned long arg, struct Qdisc *new,
1168 		     struct Qdisc **old)
1169 {
1170 	struct htb_class *cl = (struct htb_class *)arg;
1171 
1172 	if (cl->level)
1173 		return -EINVAL;
1174 	if (new == NULL &&
1175 	    (new = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1176 				     cl->common.classid)) == NULL)
1177 		return -ENOBUFS;
1178 
1179 	*old = qdisc_replace(sch, new, &cl->un.leaf.q);
1180 	return 0;
1181 }
1182 
1183 static struct Qdisc *htb_leaf(struct Qdisc *sch, unsigned long arg)
1184 {
1185 	struct htb_class *cl = (struct htb_class *)arg;
1186 	return !cl->level ? cl->un.leaf.q : NULL;
1187 }
1188 
1189 static void htb_qlen_notify(struct Qdisc *sch, unsigned long arg)
1190 {
1191 	struct htb_class *cl = (struct htb_class *)arg;
1192 
1193 	htb_deactivate(qdisc_priv(sch), cl);
1194 }
1195 
1196 static inline int htb_parent_last_child(struct htb_class *cl)
1197 {
1198 	if (!cl->parent)
1199 		/* the root class */
1200 		return 0;
1201 	if (cl->parent->children > 1)
1202 		/* not the last child */
1203 		return 0;
1204 	return 1;
1205 }
1206 
1207 static void htb_parent_to_leaf(struct htb_sched *q, struct htb_class *cl,
1208 			       struct Qdisc *new_q)
1209 {
1210 	struct htb_class *parent = cl->parent;
1211 
1212 	WARN_ON(cl->level || !cl->un.leaf.q || cl->prio_activity);
1213 
1214 	if (parent->cmode != HTB_CAN_SEND)
1215 		htb_safe_rb_erase(&parent->pq_node,
1216 				  &q->hlevel[parent->level].wait_pq);
1217 
1218 	parent->level = 0;
1219 	memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1220 	INIT_LIST_HEAD(&parent->un.leaf.drop_list);
1221 	parent->un.leaf.q = new_q ? new_q : &noop_qdisc;
1222 	parent->tokens = parent->buffer;
1223 	parent->ctokens = parent->cbuffer;
1224 	parent->t_c = ktime_get_ns();
1225 	parent->cmode = HTB_CAN_SEND;
1226 }
1227 
1228 static void htb_destroy_class(struct Qdisc *sch, struct htb_class *cl)
1229 {
1230 	if (!cl->level) {
1231 		WARN_ON(!cl->un.leaf.q);
1232 		qdisc_destroy(cl->un.leaf.q);
1233 	}
1234 	gen_kill_estimator(&cl->rate_est);
1235 	tcf_block_put(cl->block);
1236 	kfree(cl);
1237 }
1238 
1239 static void htb_destroy(struct Qdisc *sch)
1240 {
1241 	struct htb_sched *q = qdisc_priv(sch);
1242 	struct hlist_node *next;
1243 	struct htb_class *cl;
1244 	unsigned int i;
1245 
1246 	cancel_work_sync(&q->work);
1247 	qdisc_watchdog_cancel(&q->watchdog);
1248 	/* This line used to be after htb_destroy_class call below
1249 	 * and surprisingly it worked in 2.4. But it must precede it
1250 	 * because filter need its target class alive to be able to call
1251 	 * unbind_filter on it (without Oops).
1252 	 */
1253 	tcf_block_put(q->block);
1254 
1255 	for (i = 0; i < q->clhash.hashsize; i++) {
1256 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1257 			tcf_block_put(cl->block);
1258 			cl->block = NULL;
1259 		}
1260 	}
1261 	for (i = 0; i < q->clhash.hashsize; i++) {
1262 		hlist_for_each_entry_safe(cl, next, &q->clhash.hash[i],
1263 					  common.hnode)
1264 			htb_destroy_class(sch, cl);
1265 	}
1266 	qdisc_class_hash_destroy(&q->clhash);
1267 	__qdisc_reset_queue(&q->direct_queue);
1268 }
1269 
1270 static int htb_delete(struct Qdisc *sch, unsigned long arg)
1271 {
1272 	struct htb_sched *q = qdisc_priv(sch);
1273 	struct htb_class *cl = (struct htb_class *)arg;
1274 	struct Qdisc *new_q = NULL;
1275 	int last_child = 0;
1276 
1277 	/* TODO: why don't allow to delete subtree ? references ? does
1278 	 * tc subsys guarantee us that in htb_destroy it holds no class
1279 	 * refs so that we can remove children safely there ?
1280 	 */
1281 	if (cl->children || cl->filter_cnt)
1282 		return -EBUSY;
1283 
1284 	if (!cl->level && htb_parent_last_child(cl)) {
1285 		new_q = qdisc_create_dflt(sch->dev_queue, &pfifo_qdisc_ops,
1286 					  cl->parent->common.classid);
1287 		last_child = 1;
1288 	}
1289 
1290 	sch_tree_lock(sch);
1291 
1292 	if (!cl->level) {
1293 		unsigned int qlen = cl->un.leaf.q->q.qlen;
1294 		unsigned int backlog = cl->un.leaf.q->qstats.backlog;
1295 
1296 		qdisc_reset(cl->un.leaf.q);
1297 		qdisc_tree_reduce_backlog(cl->un.leaf.q, qlen, backlog);
1298 	}
1299 
1300 	/* delete from hash and active; remainder in destroy_class */
1301 	qdisc_class_hash_remove(&q->clhash, &cl->common);
1302 	if (cl->parent)
1303 		cl->parent->children--;
1304 
1305 	if (cl->prio_activity)
1306 		htb_deactivate(q, cl);
1307 
1308 	if (cl->cmode != HTB_CAN_SEND)
1309 		htb_safe_rb_erase(&cl->pq_node,
1310 				  &q->hlevel[cl->level].wait_pq);
1311 
1312 	if (last_child)
1313 		htb_parent_to_leaf(q, cl, new_q);
1314 
1315 	sch_tree_unlock(sch);
1316 
1317 	htb_destroy_class(sch, cl);
1318 	return 0;
1319 }
1320 
1321 static int htb_change_class(struct Qdisc *sch, u32 classid,
1322 			    u32 parentid, struct nlattr **tca,
1323 			    unsigned long *arg)
1324 {
1325 	int err = -EINVAL;
1326 	struct htb_sched *q = qdisc_priv(sch);
1327 	struct htb_class *cl = (struct htb_class *)*arg, *parent;
1328 	struct nlattr *opt = tca[TCA_OPTIONS];
1329 	struct nlattr *tb[TCA_HTB_MAX + 1];
1330 	struct tc_htb_opt *hopt;
1331 	u64 rate64, ceil64;
1332 
1333 	/* extract all subattrs from opt attr */
1334 	if (!opt)
1335 		goto failure;
1336 
1337 	err = nla_parse_nested(tb, TCA_HTB_MAX, opt, htb_policy, NULL);
1338 	if (err < 0)
1339 		goto failure;
1340 
1341 	err = -EINVAL;
1342 	if (tb[TCA_HTB_PARMS] == NULL)
1343 		goto failure;
1344 
1345 	parent = parentid == TC_H_ROOT ? NULL : htb_find(parentid, sch);
1346 
1347 	hopt = nla_data(tb[TCA_HTB_PARMS]);
1348 	if (!hopt->rate.rate || !hopt->ceil.rate)
1349 		goto failure;
1350 
1351 	/* Keeping backward compatible with rate_table based iproute2 tc */
1352 	if (hopt->rate.linklayer == TC_LINKLAYER_UNAWARE)
1353 		qdisc_put_rtab(qdisc_get_rtab(&hopt->rate, tb[TCA_HTB_RTAB]));
1354 
1355 	if (hopt->ceil.linklayer == TC_LINKLAYER_UNAWARE)
1356 		qdisc_put_rtab(qdisc_get_rtab(&hopt->ceil, tb[TCA_HTB_CTAB]));
1357 
1358 	if (!cl) {		/* new class */
1359 		struct Qdisc *new_q;
1360 		int prio;
1361 		struct {
1362 			struct nlattr		nla;
1363 			struct gnet_estimator	opt;
1364 		} est = {
1365 			.nla = {
1366 				.nla_len	= nla_attr_size(sizeof(est.opt)),
1367 				.nla_type	= TCA_RATE,
1368 			},
1369 			.opt = {
1370 				/* 4s interval, 16s averaging constant */
1371 				.interval	= 2,
1372 				.ewma_log	= 2,
1373 			},
1374 		};
1375 
1376 		/* check for valid classid */
1377 		if (!classid || TC_H_MAJ(classid ^ sch->handle) ||
1378 		    htb_find(classid, sch))
1379 			goto failure;
1380 
1381 		/* check maximal depth */
1382 		if (parent && parent->parent && parent->parent->level < 2) {
1383 			pr_err("htb: tree is too deep\n");
1384 			goto failure;
1385 		}
1386 		err = -ENOBUFS;
1387 		cl = kzalloc(sizeof(*cl), GFP_KERNEL);
1388 		if (!cl)
1389 			goto failure;
1390 
1391 		err = tcf_block_get(&cl->block, &cl->filter_list);
1392 		if (err) {
1393 			kfree(cl);
1394 			goto failure;
1395 		}
1396 		if (htb_rate_est || tca[TCA_RATE]) {
1397 			err = gen_new_estimator(&cl->bstats, NULL,
1398 						&cl->rate_est,
1399 						NULL,
1400 						qdisc_root_sleeping_running(sch),
1401 						tca[TCA_RATE] ? : &est.nla);
1402 			if (err) {
1403 				tcf_block_put(cl->block);
1404 				kfree(cl);
1405 				goto failure;
1406 			}
1407 		}
1408 
1409 		cl->children = 0;
1410 		INIT_LIST_HEAD(&cl->un.leaf.drop_list);
1411 		RB_CLEAR_NODE(&cl->pq_node);
1412 
1413 		for (prio = 0; prio < TC_HTB_NUMPRIO; prio++)
1414 			RB_CLEAR_NODE(&cl->node[prio]);
1415 
1416 		/* create leaf qdisc early because it uses kmalloc(GFP_KERNEL)
1417 		 * so that can't be used inside of sch_tree_lock
1418 		 * -- thanks to Karlis Peisenieks
1419 		 */
1420 		new_q = qdisc_create_dflt(sch->dev_queue,
1421 					  &pfifo_qdisc_ops, classid);
1422 		sch_tree_lock(sch);
1423 		if (parent && !parent->level) {
1424 			unsigned int qlen = parent->un.leaf.q->q.qlen;
1425 			unsigned int backlog = parent->un.leaf.q->qstats.backlog;
1426 
1427 			/* turn parent into inner node */
1428 			qdisc_reset(parent->un.leaf.q);
1429 			qdisc_tree_reduce_backlog(parent->un.leaf.q, qlen, backlog);
1430 			qdisc_destroy(parent->un.leaf.q);
1431 			if (parent->prio_activity)
1432 				htb_deactivate(q, parent);
1433 
1434 			/* remove from evt list because of level change */
1435 			if (parent->cmode != HTB_CAN_SEND) {
1436 				htb_safe_rb_erase(&parent->pq_node, &q->hlevel[0].wait_pq);
1437 				parent->cmode = HTB_CAN_SEND;
1438 			}
1439 			parent->level = (parent->parent ? parent->parent->level
1440 					 : TC_HTB_MAXDEPTH) - 1;
1441 			memset(&parent->un.inner, 0, sizeof(parent->un.inner));
1442 		}
1443 		/* leaf (we) needs elementary qdisc */
1444 		cl->un.leaf.q = new_q ? new_q : &noop_qdisc;
1445 
1446 		cl->common.classid = classid;
1447 		cl->parent = parent;
1448 
1449 		/* set class to be in HTB_CAN_SEND state */
1450 		cl->tokens = PSCHED_TICKS2NS(hopt->buffer);
1451 		cl->ctokens = PSCHED_TICKS2NS(hopt->cbuffer);
1452 		cl->mbuffer = 60ULL * NSEC_PER_SEC;	/* 1min */
1453 		cl->t_c = ktime_get_ns();
1454 		cl->cmode = HTB_CAN_SEND;
1455 
1456 		/* attach to the hash list and parent's family */
1457 		qdisc_class_hash_insert(&q->clhash, &cl->common);
1458 		if (parent)
1459 			parent->children++;
1460 		if (cl->un.leaf.q != &noop_qdisc)
1461 			qdisc_hash_add(cl->un.leaf.q, true);
1462 	} else {
1463 		if (tca[TCA_RATE]) {
1464 			err = gen_replace_estimator(&cl->bstats, NULL,
1465 						    &cl->rate_est,
1466 						    NULL,
1467 						    qdisc_root_sleeping_running(sch),
1468 						    tca[TCA_RATE]);
1469 			if (err)
1470 				return err;
1471 		}
1472 		sch_tree_lock(sch);
1473 	}
1474 
1475 	rate64 = tb[TCA_HTB_RATE64] ? nla_get_u64(tb[TCA_HTB_RATE64]) : 0;
1476 
1477 	ceil64 = tb[TCA_HTB_CEIL64] ? nla_get_u64(tb[TCA_HTB_CEIL64]) : 0;
1478 
1479 	psched_ratecfg_precompute(&cl->rate, &hopt->rate, rate64);
1480 	psched_ratecfg_precompute(&cl->ceil, &hopt->ceil, ceil64);
1481 
1482 	/* it used to be a nasty bug here, we have to check that node
1483 	 * is really leaf before changing cl->un.leaf !
1484 	 */
1485 	if (!cl->level) {
1486 		u64 quantum = cl->rate.rate_bytes_ps;
1487 
1488 		do_div(quantum, q->rate2quantum);
1489 		cl->quantum = min_t(u64, quantum, INT_MAX);
1490 
1491 		if (!hopt->quantum && cl->quantum < 1000) {
1492 			pr_warn("HTB: quantum of class %X is small. Consider r2q change.\n",
1493 				cl->common.classid);
1494 			cl->quantum = 1000;
1495 		}
1496 		if (!hopt->quantum && cl->quantum > 200000) {
1497 			pr_warn("HTB: quantum of class %X is big. Consider r2q change.\n",
1498 				cl->common.classid);
1499 			cl->quantum = 200000;
1500 		}
1501 		if (hopt->quantum)
1502 			cl->quantum = hopt->quantum;
1503 		if ((cl->prio = hopt->prio) >= TC_HTB_NUMPRIO)
1504 			cl->prio = TC_HTB_NUMPRIO - 1;
1505 	}
1506 
1507 	cl->buffer = PSCHED_TICKS2NS(hopt->buffer);
1508 	cl->cbuffer = PSCHED_TICKS2NS(hopt->cbuffer);
1509 
1510 	sch_tree_unlock(sch);
1511 
1512 	qdisc_class_hash_grow(sch, &q->clhash);
1513 
1514 	*arg = (unsigned long)cl;
1515 	return 0;
1516 
1517 failure:
1518 	return err;
1519 }
1520 
1521 static struct tcf_block *htb_tcf_block(struct Qdisc *sch, unsigned long arg)
1522 {
1523 	struct htb_sched *q = qdisc_priv(sch);
1524 	struct htb_class *cl = (struct htb_class *)arg;
1525 
1526 	return cl ? cl->block : q->block;
1527 }
1528 
1529 static unsigned long htb_bind_filter(struct Qdisc *sch, unsigned long parent,
1530 				     u32 classid)
1531 {
1532 	struct htb_class *cl = htb_find(classid, sch);
1533 
1534 	/*if (cl && !cl->level) return 0;
1535 	 * The line above used to be there to prevent attaching filters to
1536 	 * leaves. But at least tc_index filter uses this just to get class
1537 	 * for other reasons so that we have to allow for it.
1538 	 * ----
1539 	 * 19.6.2002 As Werner explained it is ok - bind filter is just
1540 	 * another way to "lock" the class - unlike "get" this lock can
1541 	 * be broken by class during destroy IIUC.
1542 	 */
1543 	if (cl)
1544 		cl->filter_cnt++;
1545 	return (unsigned long)cl;
1546 }
1547 
1548 static void htb_unbind_filter(struct Qdisc *sch, unsigned long arg)
1549 {
1550 	struct htb_class *cl = (struct htb_class *)arg;
1551 
1552 	if (cl)
1553 		cl->filter_cnt--;
1554 }
1555 
1556 static void htb_walk(struct Qdisc *sch, struct qdisc_walker *arg)
1557 {
1558 	struct htb_sched *q = qdisc_priv(sch);
1559 	struct htb_class *cl;
1560 	unsigned int i;
1561 
1562 	if (arg->stop)
1563 		return;
1564 
1565 	for (i = 0; i < q->clhash.hashsize; i++) {
1566 		hlist_for_each_entry(cl, &q->clhash.hash[i], common.hnode) {
1567 			if (arg->count < arg->skip) {
1568 				arg->count++;
1569 				continue;
1570 			}
1571 			if (arg->fn(sch, (unsigned long)cl, arg) < 0) {
1572 				arg->stop = 1;
1573 				return;
1574 			}
1575 			arg->count++;
1576 		}
1577 	}
1578 }
1579 
1580 static const struct Qdisc_class_ops htb_class_ops = {
1581 	.graft		=	htb_graft,
1582 	.leaf		=	htb_leaf,
1583 	.qlen_notify	=	htb_qlen_notify,
1584 	.find		=	htb_search,
1585 	.change		=	htb_change_class,
1586 	.delete		=	htb_delete,
1587 	.walk		=	htb_walk,
1588 	.tcf_block	=	htb_tcf_block,
1589 	.bind_tcf	=	htb_bind_filter,
1590 	.unbind_tcf	=	htb_unbind_filter,
1591 	.dump		=	htb_dump_class,
1592 	.dump_stats	=	htb_dump_class_stats,
1593 };
1594 
1595 static struct Qdisc_ops htb_qdisc_ops __read_mostly = {
1596 	.cl_ops		=	&htb_class_ops,
1597 	.id		=	"htb",
1598 	.priv_size	=	sizeof(struct htb_sched),
1599 	.enqueue	=	htb_enqueue,
1600 	.dequeue	=	htb_dequeue,
1601 	.peek		=	qdisc_peek_dequeued,
1602 	.init		=	htb_init,
1603 	.reset		=	htb_reset,
1604 	.destroy	=	htb_destroy,
1605 	.dump		=	htb_dump,
1606 	.owner		=	THIS_MODULE,
1607 };
1608 
1609 static int __init htb_module_init(void)
1610 {
1611 	return register_qdisc(&htb_qdisc_ops);
1612 }
1613 static void __exit htb_module_exit(void)
1614 {
1615 	unregister_qdisc(&htb_qdisc_ops);
1616 }
1617 
1618 module_init(htb_module_init)
1619 module_exit(htb_module_exit)
1620 MODULE_LICENSE("GPL");
1621